Benchmarking and improving imputation approaches for recurrent inversions in the human genome

Inversions are a type of structural variant that are involved in phenotypic differences among individuals. Due to certain features, such as the fact of usually implying no loss or gain of DNA or the presence of large inverted repeats at their breakpoints, the characterization of inversions is quite difficult. It has been recently found that in humans many inversions are recurrent and they are not linked to other genomic variants. For this reason, the effect of recurrent inversions has been largely missed in current GWAS, and it is necessary to develop new methods to predict inversion genotypes accurately in the datasets of interest. Here, we have done a benchmarking analysis of the genotype predictions among different imputation tools, comparing IMPUTE2 with other softwares, such as IMPUTE5, BEAGLE or/and scoreInvHap. The accuracy was calculated as the r 2  between experimental and imputed genotypes. From our set of 130 experimentally genotyped inversions 55 are recurrent. We found out that 23 and 18 out of 55 were imputable (r 2  > 0.8) in European and African populations, respectively, but it varies among softwares. Nevertheless, this ratio increases to 26/55 for both populations when we filter out samples with a post-imputation genotype probability lower than 0.8. Finally, we are also testing a tool based on deep learning which could avoid the HMM-based algorithm limitations, as the position or linkage dependence and region complexity, to increase our catalogue of imputable inversions.